| Background: Acute ischemic stroke(AIS)is a disease with high morbidity,disability,mortality,and recurrence rates,and early revascularization and reperfusion are associated with a good clinical prognosis.Mechanical thrombectomy(MTB)is the first-line treatment for patients with acute ischemic stroke combined with large vessel occlusion,and several endovascular treatment strategies are available.However,there are no quantitative methods to select the best endovascular treatment strategy or to predict the difficulty of thrombus clearance.Radiomics is an emerging field aimed at high-throughput extraction of quantitative,reproducible,hard-to-identify by the human eye,quantitative information from imaging images.These high-dimensional features exhibit superior diagnostic advantages and capabilities,good sensitivity,and specificity in predicting prognosis and treatment strategies and may in the future be a predictor of successful recanalization by mechanical thrombus removal.Objective: Our study is based on the radiomics of susceptibility-weighted imaging(SWI)to discover radiomic features(RF)that are difficult for the human eye to identify and quantify.We used RF extracted from SWI thrombus to develop and validate a predictive model: identifying patients who achieved first recanalization using thrombus aspiration.The diagnostic power and stability of the model were assessed and the value of this technique in predicting the surgical strategy for MTB in patients with AIS was explored.Methods: This study consecutively included patients with acute anterior circulation large vessel(middle cerebral artery M1 and M2 segments)occlusive cerebral infarction who underwent mechanical thrombectomy in the Department of Neurology at the Third Affiliated Hospital of Northwestern University,Xi’an,China,between January 2018 and February2022.We included patients with large vessel occlusion-AIS(LVO-AIS)and treated with MTB within 24 hours after the onset of the disease.The preoperative SWI images were imported into 3D Slicer(version 4.11;Harvard University,National Institutes of Health)software in DICOM format to manually segment the regions of interest(ROI)[26].In addition,we viewed the corresponding magnetic resonance angiography(MRA)images for guidance.The segmented regions of interest were based on the susceptibility vessel sign(SVS).Interobserver reproducibility of lesion segmentation was assessed by calculating the kappa coefficient of the extracted RFs.Good agreement was defined as RFs with a kappa coefficient greater than 0.75.Based on the thrombus regions present in the SVS sign,1874 RFs were automatically extracted from the thrombus using pyradiomics.prior to building the model,we used the Mann-Whitney U test to initially screen the training cohort of all 1874 previously extracted RFs for those significantly associated with the first attempt at recanalization after thrombus aspiration(p<0.01).We used the least absolute shrinkage and selection operator(LASSO)regression model for feature dimensionality reduction and placed the initially selected features into the model,allowing the weights of insignificant features to be set to 0.The predictive performance of the radiomics models was evaluated using accuracy,specificity,sensitivity,negative predictive value,positive predictive value,and area under the receiver operating characteristic curve(AUC).RF selection and ML model development were performed on the open-source Python library SCRICIT-LEARING(version 0.21.3).The included patients were divided into two groups according to the success of catheter aspiration,and the normally distributed measures were expressed as mean±standard deviation using the independent samples t-test;the measures that did not conform to the normal distribution were expressed as median and interquartile spacing,and the Wilcoxon rank sum test was used for comparison;the count data were expressed as frequencies and percentages,and the χ2 test or Fisher’s exact probability method was used for comparison between groups.p< 0.05 was considered a statistically significant difference.Result:(1)The study ultimately included 136 participants,with 102 in the training cohort and 34 in the test cohort.There was no statistical significance between the two groups in terms of age,sex,previous stroke,NIHSS,right-sided occlusion,occluded vessels,hypertension,diabetes,AF,smoking and HCY,hyperlipidemia,and IV-rtpa(p> 0.05).There were 87 cases in the No FAR(First-attempt recanalization with a direct aspiration first pass technique,FAR)group and 49 cases in the FAR group,in both groups except for two indicators of IV-rtpa(p = 0.01)and hyperlipidemia(p = 0.02).All the rest were not statistically different from the outcome FAR(p> 0.05).(2)After applying filters to the SWI images,1874 RFs were automatically extracted from the thrombus using pyradiomics.we obtained 69 features after initial screening of the1874 RFs using the Mann-Whitney U test.The screened RFs with p < 0.01 were further put into the LASSO regression model,and as the parameter lambda increased,the regression coefficients(i.e.,vertical coordinate values)converged,eventually converging to 0.At a lambda value of 0.036,we screened the eight best imaging histology features with non-zero coefficients into the prediction model.(3)The following are the eight characteristics that were finally filtered out :original＿ngtdm＿Contrast;wavelet-HLH＿glcm＿Imc2;wavelet-HHL＿glcm＿MCC;squareroot＿ngtdm＿Contrast;original＿glcm＿Imc1;original＿glrlm＿Long Run Emphasis;wavelet-LLH＿glszm＿Small Area Emphasis;wavelet-LLH＿glszm＿Small Area High Gray Level Emphasis.(4)In the training cohort,the AUC(Area under the receiver operating characteristic curve,AUC)of the radiology model was 0.798(95% CI 0.707-0.871).In the test cohort,the AUC of the radiology model was 0.784(95% CI: 0.610-0.906).In the test cohort,the model had a predictive accuracy of 76.5%,specificity of 95.5%,sensitivity of 41.7%,positive predictive value of 83.3%,and negative predictive value of 75.0%.In the training cohort,the model had a prediction accuracy of 75.5%,specificity of 86.2%,sensitivity of 56.8%,positive predictive value of 70.0%,and negative predictive value of 77.8%.Conclusion:(1)Imaging histologic features extracted from thrombus on pretreatment SWI images may provide information about the success or difficulty of different MTB strategies in AIS patients.Predictive modeling by combining SVS signs of thrombus,MTB surgical strategy,and radiomics may be a prospective technique to personalize the mechanical thrombectomy strategy in AIS patients.The prediction model developed in this paper has high specificity but low sensitivity.(2)Finally,we identified eight imaging histological features associated with FAR.four features,Imc2,MCC,original＿ngtdm＿Contrast,and squareroot＿ngtdm＿Contrast,were positively associated with FAR outcomes.four features,Imc1,LRE,SAE,and SAHGLE,were negatively correlated.SWI techniques can provide imaging evidence for predicting first recanalization outcomes. |
